GreenCluster is two NVIDIA GB10 servers (128 GB of unified memory each) working as a duo, orchestrated by k3s + ArgoCD (GitOps). Storage lives on ZFS, replicated by Syncoid. Every piece of infrastructure is versioned in git : nothing is "hacked by hand", everything redeploys in one gesture.
Two servers, one cluster
- msi-living — the control plane (k3s controller + ArgoCD) and primary ZFS storage. An accepted SPOF, but replaceable by design.
- dell-living — the compute node (worker). Its lifecycle is automated : drain on shutdown, rejoin + cleanup on boot.
🏠 Home Assistant — the energy brain
Home Assistant is not just a detail : it is what orchestrates the cluster's energy use. When msi's unified memory drops below a critical threshold (< 11 %), HA decides whether to run the big lifeline or flush it.
- The 3-state machine —
solo(light model on msi) /stop(flush) /cluster(DeepSeek on both). - The MQTT sensor
home/greencluster/modepublishes the current mode = source of truth. - The memory gate — HA only starts
clusterif memory is available : the lifeline can never crash again. - Power — HA even powers dell-living on and off as needed.
Learn more about the hero watching over us : home-assistant.io.
✨ The lifeline — DeepSeek on vLLM
The heart that keeps the team going : a DeepSeek V4 Flash 0731 model — see DeepSeek-V4-Flash-0731 on Hugging Face — serving an OpenAI-compatible API (alias gpt-4o) across the cluster, powered by vLLM (the open-source inference server). It runs thanks to a community Docker image built for the NVIDIA GB10 : eugr/spark-vllm-docker — a kit that makes running vLLM on one or more DGX Spark easy, with ready recipes (DeepSeek, Qwen…), multi-node loading and RoCE acceleration.
- solo — a light model (Qwen) on msi, for everyday life without effort.
- cluster — DeepSeek in tensor parallelism 2 (both GB10 linked by QSFP/RoCE), when real power is needed.
The founding idea : serious AI power only when necessary, while limiting consumption the rest of the time. Sober by design 🌿.